Finally official Lc0 v0.31.0-rc1

Discussion of anything and everything relating to chess playing software and machines.

Moderators: hgm, Rebel, chrisw

Hai
Posts: 613
Joined: Sun Aug 04, 2013 1:19 pm

Re: Finally official Lc0 v0.31.0-rc1

Post by Hai »

chesskobra wrote: Wed Mar 27, 2024 2:36 pm What network should I use on an i7, 13 gen, desktop with 32 GB RAM?
BT4.
I'm using it also on my MacBook.
chesskobra
Posts: 268
Joined: Thu Jul 21, 2022 12:30 am
Full name: Chesskobra

Re: Finally official Lc0 v0.31.0-rc1

Post by chesskobra »

On this link https://lczero.org/dev/wiki/best-nets-for-lc0/ there is a BT4 network BT4-1024x15x32h-swa-6147500 - is this what you recommend? The link says it is for GPU. But if I am not mistaken my i7 13 gen (Dell XPS 8960) has only CPU. So is it still a good network?
Dann Corbit
Posts: 12662
Joined: Wed Mar 08, 2006 8:57 pm
Location: Redmond, WA USA

Re: Finally official Lc0 v0.31.0-rc1

Post by Dann Corbit »

Lc0 is lame without a GPU.
I would not bother.
Taking ideas is not a vice, it is a virtue. We have another word for this. It is called learning.
But sharing ideas is an even greater virtue. We have another word for this. It is called teaching.
Dann Corbit
Posts: 12662
Joined: Wed Mar 08, 2006 8:57 pm
Location: Redmond, WA USA

Re: Finally official Lc0 v0.31.0-rc1

Post by Dann Corbit »

Marek Soszynski wrote: Wed Mar 27, 2024 1:01 pm
Dann Corbit wrote: Wed Mar 27, 2024 11:49 am
<snip>

Here are the results for the new Lc0 (which was notoriously poor at tactics):

<snip>
Many thanks.

So, what is the comparison to the old Lc0? In your own words.
This is the result for Lc0-v0.30.0
I used network :
C:\Users\dannc\chess\lc0\v0.30.0\t2-768x15x24h-swa-5230000.pb.gz
10/13/2023 09:01 AM 168,247,318 t2-768x15x24h-swa-5230000.pb.gz
bottom line: 73 of 120 matching moves

Code: Select all

Analysis of H:\hardest.epd   
Analyzing engine: Lc0-v0.30.0
3/27/2024 4:07:27 AM Level: 60 Seconds
   Source                    : List of EPD/PGN files
      H:\hardest.epd
   Colours                   : White moves=True, Black moves=True
   Direction                 : forward
   Games (PGN)               : From 1 to 999999
   Moves                     : All moves
   Positions (EPD)           : All positions
   Use Engine(s)             : List
      Lc0-v0.30.0
      Stockfish_24032617_x64_avx2
   Use seconds per move      : 60 Seconds per move
   Analysis Lines            : Minimum search depth=2
   Activate abort analysis   : True. Minimal analysis depth 9. Abort, if the correct move is held 220 additional ply
  1) g5-g6              Kb1-b2   
  2) b5-b6              Be2-f1   
  3) Kb8-a7             Kb8-a7  * 0 Seconds
  4) .. Rc5-c2          Kg7-g6   
  5) .. Nc5-d7          Nc5-d7  * 18 Seconds
  6) Qa3-f3             Nc5-d3   
  7) Re1-e4             Re1-e4  * 0 Seconds
  8) a6-a7              a6-a7   * 11 Seconds
  9) Bc1-b2             Bc1-b2  * 3 Seconds
 10) Ne1-d3, Ne1-f3     Ne1-f3  * 0 Seconds
 11) Ra4-a8             h7-h8Q   
 12) Be3-c1             Kd7-e6   
 13) Rd1-b1             Rd1-d3   
 14) h6-h7              c6-c7    
 15) .. e5-e4           e5-e4   * 8 Seconds
 16) Qa5xd5             Qa5xd5  * 1 Second
 17) Qc8xh8             Qc8xh8  * 5 Seconds
 18) .. Bf8-e7          Bf8-e7  * 7 Seconds
 19) .. Bb6xd4          Qf6xh4   
 20) .. Qc5-d4          Qc5-c4   
 21) .. Be2-h5          Be2-h5  * 57 Seconds
 22) Rd6-d8             Rd6-d8  * 20 Seconds
 23) Qg6-f5             Qg6-f7   
 24) Rc4-c7             Rc4-c7  * 17 Seconds
 25) Rg3-g5             Rg3-g5  * 19 Seconds
 26) Nd4-f5             Qe2-h5   
 27) Re3-c3             Re3-c3  * 30 Seconds
 28) f2-f4              f2-f4   * 0 Seconds
 29) Bc8xa6             Bc8xa6  * 3 Seconds
 30) Bc8xa6             d2-d4    
 31) Kc8-c7             Qd2-d8   
 32) Ne7-g6             Ne7-g6  * 0 Seconds
 33) Kc8-d7, Kc8-d8     Kc8-d7  * 0 Seconds
 34) Nc8-e7             Nc8-d6   
 35) Nc8-e7             Qe3-d4   
 36) Rc8-a8             Rc8-a8  * 0 Seconds
 37) b2-b3              b2-b3   * 33 Seconds
 38) Be1-h4             Be1-h4  * 8 Seconds
 39) Bb6-a7             Bb6-a7  * 0 Seconds
 40) d6-d7              d6-d7   * 0 Seconds
 41) Qa7xc7             Qa7-c5   
 42) Ne4-d6             Ne4-d6  * 3 Seconds
 43) Ng6-h4             Rg3xg4   
 44) Rg8-h8             g3-g4    
 45) Bd4xg7             Bd4xg7  * 1 Second
 46) Kd5-e6             Qd4-f2   
 47) Kg4-f3             Kg4-f3  * 1 Second
 48) Qh5xg6             Bh4-f6   
 49) Rc4xc5             Rc4xc5  * 5 Seconds
 50) .. h7-h5           h7-h5   * 43 Seconds
 51) Nd4-f3             Rg6xa6   
 52) .. Bb7-f3          Bb7-f3  * 6 Seconds
 53) Rf1-f5             Rf1-f5  * 0 Seconds
 54) e7-e8Q, e7-e8R     e7-e8Q  * 46 Seconds
 55) Be5-b8             Be5-d4   
 56) g4-g5              g4-g5   * 3 Seconds
 57) Ra5-a6             Nd5-c3   
 58) Ke2-f3, Bc2-b1     Ke2-f3  * 14 Seconds
 59) a3-a4              h4-h5    
 60) Be4-d5             Be4-d5  * 0 Seconds
 61) .. Rd8-d6, b7-b6   Rd8-d6  * 2 Seconds
 62) Rd2-d3             Rd2-d3  * 56 Seconds
 63) Qc4xb5             Qc4xb5  * 3 Seconds
 64) Bb7-c6             Nd1-b2   
 65) Rf2-f1             Rf2-f1  * 7 Seconds
 66) Be4-h1             Be4-f3   
 67) Nh4xg6             Nh4xg6  * 0 Seconds
 68) .. Ng2xe3, Nd5xe3  Ng2xe3  * 0 Seconds
 69) c4xd5              c4-c5    
 70) h6-h7              Qh3-c3   
 71) Nf4-h3             Nf4xg2   
 72) c3-c4              Kh1-g1   
 73) Bg2xb7             Qc4-h4   
 74) Qc7-f4             Qc7xc5   
 75) .. h6-h5           Ne3xg2   
 76) Qb1-b6             Qb1-c1   
 77) Nc4-e5             Nc4xb6   
 78) Kf8-e7, Kf8-e8     Kf8-e7  * 1 Second
 79) Rg8-h8             Rg8-h8  * 16 Seconds
 80) Rf6-h6             Rf6-f7   
 81) Rg6-g4             Nf8-e6   
 82) Rb3-c3             Be2xb5   
 83) Rh4-a4             Rh4-a4  * 0 Seconds
 84) Nb5-d4             Nb5-d4  * 12 Seconds
 85) c2-c3              c2-c3   * 22 Seconds
 86) Qg3-g8             Qg3-g8  * 0 Seconds
 87) Kh4-g5             Kh4-g5  * 3 Seconds
 88) .. Qd1-e1          Qd1-e1  * 4 Seconds
 89) Bc1-b2             Bc1-b2  * 0 Seconds
 90) .. Rg5xg2, Ne1xg2  Rg5xg2  * 0 Seconds
 91) Qc5xf8             Qc5xf8  * 0 Seconds
 92) Qf1-g1             Qf1-g1  * 12 Seconds
 93) Bb7-d5             Na7-b5   
 94) Qe4-e5, Qe4-c6, Qe4-b7, Kg2-g1  Kg2-g1  * 4 Seconds
 95) Nc7-a8             Nc7-e8   
 96) Ne2-g1             Ne2-g1  * 48 Seconds
 97) Bc1-d2             Bc1-d2  * 39 Seconds
 98) b2-b3              Nd5-f6   
 99) Nh4-g6             Nh4-g6  * 0 Seconds
100) .. Kh8-h7, Kh8-g7  Kh8-h7  * 2 Seconds
101) Bg3-f4             Bg3-f4  * 20 Seconds
102) .. Rc1-d1          Rc1-d1  * 0 Seconds
103) Bf4-h6             Bf4-h6  * 20 Seconds
104) e5-e6              e5xf6    
105) Ke5-e6, Ke5-f6     Ke5-e6  * 25 Seconds
106) .. Nb5xd4          Nb5xd4  * 46 Seconds
107) Re3xd3             Re3xd3  * 8 Seconds
108) Be2-c4             Be2-c4  * 3 Seconds
109) Ne3-f1             Ne3-f5   
110) .. Rg7-h7          Rg7-h7  * 0 Seconds
111) Na3-c2             Na3-c2  * 23 Seconds
112) Ke6-f7             Be7-f6   
113) Qh2-h6             Qh2-h6  * 0 Seconds
114) Rg7-h7             Rg7-d7   
115) Bf2-g1             Bf2-g1  * 2 Seconds
116) g2-g3              g2-g3   * 0 Seconds
117) Ba7-b8             Ba7-b8  * 3 Seconds
118) Bc4-g8             Bc4-g8  * 2 Seconds
119) Rc6-c8             e6xf7    
120) .. Nf6xh5          Nf6xh5  * 18 Seconds
73 of 120 matching moves
3/27/2024 6:09:28 AM, Total time: 2:02:01 AM Rated time: 59:43 = 3583 Seconds
Just for fun, here is the bleeding edge Stockfish, which gets 94 of 120 matching moves:

Code: Select all

Analysis of H:\hardest.epd   
Analyzing engine: Stockfish_24032617_x64_avx2
3/27/2024 7:08:42 AM Level: 60 Seconds
   Source                    : List of EPD/PGN files
      H:\hardest.epd
   Colours                   : White moves=True, Black moves=True
   Direction                 : forward
   Games (PGN)               : From 1 to 999999
   Moves                     : All moves
   Positions (EPD)           : All positions
   Use Engine(s)             : List
      Stockfish_24032617_x64_avx2
   Use seconds per move      : 60 Seconds per move
   Analysis Lines            : Minimum search depth=2
   Activate abort analysis   : True. Minimal analysis depth 9. Abort, if the correct move is held 220 additional ply
  1) g5-g6              Kb1-b2   
  2) b5-b6              b5-b6   * 4 Seconds
  3) Kb8-a7             Kb8-a7  * 0 Seconds
  4) .. Rc5-c2          Rc5-c2  * 3 Seconds
  5) .. Nc5-d7          Nc5-d7  * 0 Seconds
  6) Qa3-f3             Qa3-f3  * 41 Seconds
  7) Re1-e4             Nf4-e2   
  8) a6-a7              a6-a7   * 0 Seconds
  9) Bc1-b2             Bc1-b2  * 0 Seconds
 10) Ne1-d3, Ne1-f3     Ne1-d3  * 0 Seconds
 11) Ra4-a8             Ra4-a8  * 0 Seconds
 12) Be3-c1             Be3-c1  * 0 Seconds
 13) Rd1-b1             Rd1-b1  * 0 Seconds
 14) h6-h7              h6-h7   * 0 Seconds
 15) .. e5-e4           e5-e4   * 1 Second
 16) Qa5xd5             Qa5xd5  * 0 Seconds
 17) Qc8xh8             Qc8xh8  * 0 Seconds
 18) .. Bf8-e7          Qb8-c7   
 19) .. Bb6xd4          Bb6xd4  * 0 Seconds
 20) .. Qc5-d4          Qc5-d4  * 0 Seconds
 21) .. Be2-h5          Be2-h5  * 0 Seconds
 22) Rd6-d8             Rd6-d8  * 0 Seconds
 23) Qg6-f5             Qg6-f5  * 5 Seconds
 24) Rc4-c7             Rc4-c7  * 4 Seconds
 25) Rg3-g5             Rg3-g5  * 0 Seconds
 26) Nd4-f5             Nd4-f5  * 0 Seconds
 27) Re3-c3             Nh4-f5   
 28) f2-f4              f2-f4   * 0 Seconds
 29) Bc8xa6             Bc8-g4   
 30) Bc8xa6             d2-d4    
 31) Kc8-c7             Nf4-d3   
 32) Ne7-g6             Ne7-g6  * 2 Seconds
 33) Kc8-d7, Kc8-d8     Kc8xc7   
 34) Nc8-e7             Nc8-d6   
 35) Nc8-e7             Nc8-e7  * 3 Seconds
 36) Rc8-a8             Rc8-a8  * 0 Seconds
 37) b2-b3              b2-b3   * 56 Seconds
 38) Be1-h4             Be1-g3   
 39) Bb6-a7             Bb6-a7  * 0 Seconds
 40) d6-d7              d6-d7   * 0 Seconds
 41) Qa7xc7             Qa7-d4   
 42) Ne4-d6             Ne4-d6  * 0 Seconds
 43) Ng6-h4             Ng6-h4  * 6 Seconds
 44) Rg8-h8             Rg8-h8  * 2 Seconds
 45) Bd4xg7             Bd4xg7  * 6 Seconds
 46) Kd5-e6             Qd4-f2   
 47) Kg4-f3             Kg4-f3  * 0 Seconds
 48) Qh5xg6             Bh4-f6   
 49) Rc4xc5             Rc4xc5  * 0 Seconds
 50) .. h7-h5           h7-h5   * 0 Seconds
 51) Nd4-f3             Nd4-f3  * 5 Seconds
 52) .. Bb7-f3          Bb7-f3  * 0 Seconds
 53) Rf1-f5             Rf1-f5  * 0 Seconds
 54) e7-e8Q, e7-e8R     e7-e8Q  * 5 Seconds
 55) Be5-b8             Be5-b8  * 18 Seconds
 56) g4-g5              Rd8-a8   
 57) Ra5-a6             Ra5-a6  * 0 Seconds
 58) Ke2-f3, Bc2-b1     Ke2-f3  * 0 Seconds
 59) a3-a4              a3-a4   * 5 Seconds
 60) Be4-d5             Be4-d5  * 0 Seconds
 61) .. Rd8-d6, b7-b6   Rd8-d6  * 0 Seconds
 62) Rd2-d3             Rd2-d3  * 1 Second
 63) Qc4xb5             Qc4xb5  * 57 Seconds
 64) Bb7-c6             Nd1-b2   
 65) Rf2-f1             Rf2-f1  * 18 Seconds
 66) Be4-h1             Be4-f3   
 67) Nh4xg6             Nc5-d3   
 68) .. Ng2xe3, Nd5xe3  Nd5xe3  * 0 Seconds
 69) c4xd5              c4xd5   * 3 Seconds
 70) h6-h7              h6-h7   * 7 Seconds
 71) Nf4-h3             Nf4-h3  * 22 Seconds
 72) c3-c4              c3-c4   * 1 Second
 73) Bg2xb7             Bg2xb7  * 3 Seconds
 74) Qc7-f4             Qc7xc5   
 75) .. h6-h5           h6-h5   * 0 Seconds
 76) Qb1-b6             Qb1-b6  * 3 Seconds
 77) Nc4-e5             Nc4-e5  * 0 Seconds
 78) Kf8-e7, Kf8-e8     Kf8-e7  * 0 Seconds
 79) Rg8-h8             Rg8-h8  * 1 Second
 80) Rf6-h6             Rf6-f7   
 81) Rg6-g4             Nf8-e6   
 82) Rb3-c3             Rb3-c3  * 20 Seconds
 83) Rh4-a4             Rh4-a4  * 8 Seconds
 84) Nb5-d4             Nb5-d4  * 0 Seconds
 85) c2-c3              c2-c3   * 1 Second
 86) Qg3-g8             Qg3-g8  * 0 Seconds
 87) Kh4-g5             Kh4-g5  * 1 Second
 88) .. Qd1-e1          Qd1-e1  * 0 Seconds
 89) Bc1-b2             Kd1-d2   
 90) .. Rg5xg2, Ne1xg2  Rg5xg2  * 18 Seconds
 91) Qc5xf8             Qc5xf8  * 0 Seconds
 92) Qf1-g1             Qf1-g1  * 0 Seconds
 93) Bb7-d5             Bb7-d5  * 5 Seconds
 94) Qe4-e5, Qe4-c6, Qe4-b7, Kg2-g1  Kg2-g1  * 0 Seconds
 95) Nc7-a8             Nc7-a8  * 57 Seconds
 96) Ne2-g1             Ne2-g1  * 3 Seconds
 97) Bc1-d2             Bc1-d2  * 0 Seconds
 98) b2-b3              Nd5-f6   
 99) Nh4-g6             Nh4-f5   
100) .. Kh8-h7, Kh8-g7  Kh8-h7  * 0 Seconds
101) Bg3-f4             Bg3-f4  * 0 Seconds
102) .. Rc1-d1          Rc1-d1  * 0 Seconds
103) Bf4-h6             Bf4-h6  * 0 Seconds
104) e5-e6              e5-e6   * 5 Seconds
105) Ke5-e6, Ke5-f6     Ke5-e6  * 0 Seconds
106) .. Nb5xd4          Nb5xd4  * 12 Seconds
107) Re3xd3             Re3xd3  * 1 Second
108) Be2-c4             Be2-c4  * 0 Seconds
109) Ne3-f1             Ne3-f1  * 7 Seconds
110) .. Rg7-h7          Rg7-h7  * 0 Seconds
111) Na3-c2             Bh4-g5   
112) Ke6-f7             Be7-f6   
113) Qh2-h6             Qh2-h6  * 0 Seconds
114) Rg7-h7             Rg7-h7  * 0 Seconds
115) Bf2-g1             Bf2-g1  * 3 Seconds
116) g2-g3              Nb5-c7   
117) Ba7-b8             Ba7-b8  * 25 Seconds
118) Bc4-g8             Bc4-g8  * 0 Seconds
119) Rc6-c8             Rc6-c8  * 5 Seconds
120) .. Nf6xh5          Nf6xh5  * 0 Seconds
94 of 120 matching moves
3/27/2024 9:09:03 AM, Total time: 2:00:20 AM Rated time: 33:33 = 2013 Seconds
Taking ideas is not a vice, it is a virtue. We have another word for this. It is called learning.
But sharing ideas is an even greater virtue. We have another word for this. It is called teaching.
Dann Corbit
Posts: 12662
Joined: Wed Mar 08, 2006 8:57 pm
Location: Redmond, WA USA

Re: Finally official Lc0 v0.31.0-rc1

Post by Dann Corbit »

With the mega net, Lc0 outperforms SF on this test (of course we cannot read a ton into that, since I have two fast GPU cards in my machine and an AMD 3790x CPU, so my hardware is not typical.)

I will say this, I never thought that Lc0 would outshine SF on a tactical test. This new version is much better and the new net is much better too.

It is possible parameter changes account for some of it. I used some of the parameters that the Tcec version of Lc0 is using (but obviously not the memory, since my GPUs only have 8GB video ram each.)

I was not able to use some of their settings, since they did not work when I tried them. Probably because they are using a zillion dollar GPU with an ocean of video RAM.
Taking ideas is not a vice, it is a virtue. We have another word for this. It is called learning.
But sharing ideas is an even greater virtue. We have another word for this. It is called teaching.
Krzysztof Grzelak
Posts: 1565
Joined: Tue Jul 15, 2014 12:47 pm

Re: Finally official Lc0 v0.31.0-rc1

Post by Krzysztof Grzelak »

Dann Corbit wrote: Wed Mar 27, 2024 5:33 pm With the mega net, Lc0 outperforms SF on this test (of course we cannot read a ton into that, since I have two fast GPU cards in my machine and an AMD 3790x CPU, so my hardware is not typical.)

I will say this, I never thought that Lc0 would outshine SF on a tactical test. This new version is much better and the new net is much better too.

It is possible parameter changes account for some of it. I used some of the parameters that the Tcec version of Lc0 is using (but obviously not the memory, since my GPUs only have 8GB video ram each.)

I was not able to use some of their settings, since they did not work when I tried them. Probably because they are using a zillion dollar GPU with an ocean of video RAM.
Please write what settings you are writing about that you could not use.
Dann Corbit
Posts: 12662
Joined: Wed Mar 08, 2006 8:57 pm
Location: Redmond, WA USA

Re: Finally official Lc0 v0.31.0-rc1

Post by Dann Corbit »

Krzysztof Grzelak wrote: Wed Mar 27, 2024 5:38 pm
Dann Corbit wrote: Wed Mar 27, 2024 5:33 pm With the mega net, Lc0 outperforms SF on this test (of course we cannot read a ton into that, since I have two fast GPU cards in my machine and an AMD 3790x CPU, so my hardware is not typical.)

I will say this, I never thought that Lc0 would outshine SF on a tactical test. This new version is much better and the new net is much better too.

It is possible parameter changes account for some of it. I used some of the parameters that the Tcec version of Lc0 is using (but obviously not the memory, since my GPUs only have 8GB video ram each.)

I was not able to use some of their settings, since they did not work when I tried them. Probably because they are using a zillion dollar GPU with an ocean of video RAM.
Please write what settings you are writing about that you could not use.
I got the settings I wanted to try from here:
https://lczero.org/watch/competitions/tcec/

Backend and backendOptions did not work.
I also did not use the value for RamLimitMB because my my GPUs only have 8GB VRAM so I set it at 7.5 GB. If it is system RAM, then I would have set it much higher.

Is there anywhere to obtain an accurate description of each setting and what they mean?
If the answer is "browse the source" then I won't bother.
I tried reading the source to find some simple information about evaluation hashing and gave up.
https://ibb.co/YfbZNYs
Taking ideas is not a vice, it is a virtue. We have another word for this. It is called learning.
But sharing ideas is an even greater virtue. We have another word for this. It is called teaching.
Krzysztof Grzelak
Posts: 1565
Joined: Tue Jul 15, 2014 12:47 pm

Re: Finally official Lc0 v0.31.0-rc1

Post by Krzysztof Grzelak »

Dann Corbit wrote: Fri Mar 29, 2024 12:30 am I got the settings I wanted to try from here:
https://lczero.org/watch/competitions/tcec/

Backend and backendOptions did not work.
I also did not use the value for RamLimitMB because my my GPUs only have 8GB VRAM so I set it at 7.5 GB. If it is system RAM, then I would have set it much higher.

Is there anywhere to obtain an accurate description of each setting and what they mean?
If the answer is "browse the source" then I won't bother.
I tried reading the source to find some simple information about evaluation hashing and gave up.
https://ibb.co/YfbZNYs
At the very beginning, thank you for your answer. Some interesting comments on I write where you made mistakes.

Option "RamLimitMB" - it does not apply to the graphics card's RAM, but to the computer's RAM.

You don't use the engine in the pro version but in the regular version. When you use the pro version, you will notice that the number of options will increase by about 30.

There was a different version played on TCEC than you have.
Werewolf
Posts: 1899
Joined: Thu Sep 18, 2008 10:24 pm

Re: Finally official Lc0 v0.31.0-rc1

Post by Werewolf »

Hai wrote: Wed Mar 27, 2024 2:49 pm
chesskobra wrote: Wed Mar 27, 2024 2:36 pm What network should I use on an i7, 13 gen, desktop with 32 GB RAM?
BT4.
I'm using it also on my MacBook.
And how many nps does that get..?
Carbo
Posts: 3
Joined: Wed Jan 19, 2022 5:03 pm
Full name: Andreas Guettinger

Re: Finally official Lc0 v0.31.0-rc1

Post by Carbo »

Dann Corbit wrote: Fri Mar 29, 2024 12:30 am
I got the settings I wanted to try from here:
https://lczero.org/watch/competitions/tcec/

Backend and backendOptions did not work.
I also did not use the value for RamLimitMB because my my GPUs only have 8GB VRAM so I set it at 7.5 GB. If it is system RAM, then I would have set it much higher.

Is there anywhere to obtain an accurate description of each setting and what they mean?
If the answer is "browse the source" then I won't bother.
I tried reading the source to find some simple information about evaluation hashing and gave up.
https://ibb.co/YfbZNYs
Maybe this helps?

.\lc0.exe --help

Code: Select all

       --ramlimit-mb=0..100000000
               Maximum memory usage for the engine, in megabytes. The estimation is very rough,
               and can be off by a lot. For example, multiple visits to a terminal node counted
               several times, and the estimation assumes that all positions have 30 possible
               moves. When set to 0, no RAM limit is enforced.
               [UCI: RamLimitMb  DEFAULT: 0  MIN: 0  MAX: 100000000]
This is from lc0-v0.30.0-cuda, I don't have v31 installed, but should work the same.